CN-121997745-A - Intelligent sub-global basis function method for calculating electromagnetic scattering characteristics of periodic target

CN121997745ACN 121997745 ACN121997745 ACN 121997745ACN-121997745-A

Abstract

The invention discloses an intelligent sub-universe basis function method for calculating periodic target electromagnetic scattering characteristics, which comprises the steps of constructing a sub-domain network and an universe network, wherein the sub-domain network and the universe network comprise a semantic and numerical feature fusion module and a Transformer module, constructing a sub-domain and universe data set, designing a mixed data and physical loss function to train the sub-domain network and the universe network, and finally predicting current coefficients of large-scale finite periodic arrays of different unit structures by using the established model, and calculating far-field scattering based on surface currents. The invention merges the solving mode of the traditional NASED basis function method into the subdomain and the universe network. The network reasoning process replaces the filling and solving of complex and time-consuming impedance matrix equations, the subarray current coefficients of subarrays with different unit structures can be rapidly predicted by combining the sub-domain network with the migration learning, and meanwhile, the whole domain network can adapt to different unit structures without retraining.

Inventors

LU WEIBING
LIU TIAN
YANG WU
GUO FEI
SONG BINGBING

Assignees

东南大学

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (5)

1. An intelligent sub-global basis function method for calculating electromagnetic scattering characteristics of a periodic target, comprising the following steps: The method comprises the steps of 1, establishing a sub-domain network and a global network, wherein the intelligent sub-domain base function method consists of the sub-domain network and the global network, the sub-domain network is used for predicting initial current coefficients of a 3 multiplied by 3 sub-array, and the global network corrects the initial current coefficients of each unit based on global coupling of a large-scale array; Step 2, data set making, namely establishing a data set according to physical characteristics of the 3 multiplied by 3 subarrays and the large-scale array, generating a label by using a traditional NASED basis function method, and preprocessing the data; step 3, training a model, namely taking a physical equation fused with an impedance matrix equation and a numerical error equation as a loss function, dividing a training set and a testing set according to a certain proportion, training and testing a network model to obtain optimal super parameters, and finally storing optimal model parameters; And 4, calculating far-field scattering characteristics of the large-scale finite periodic structure, namely firstly solving the problem of the 3 multiplied by 3 subarray to obtain NASED basis functions, obtaining final current coefficients by utilizing a subdomain network and a universe network, and finally calculating current distribution of the finite periodic array to calculate the far-field scattering characteristics of the finite periodic structure.
2. The intelligent sub-global basis function method for calculating electromagnetic scattering characteristics of a periodic object according to claim 1, wherein the implementation process of step 1 is as follows: The method comprises the steps of 1.1, establishing a sub-domain network, wherein the sub-domain network comprises a semantic and numerical feature fusion module and a Transformer module, firstly, constructing a semantic information base by using an upper left corner, an upper right corner, a left edge, an inner center, a right edge, a lower left corner, a lower right corner and a lower right corner of an array unit in the semantic and numerical feature fusion module, then converting space semantic information of each unit into a trainable vector based on the existing Wordembedding technology, linearly mapping the semantic vector and numerical feature vectors such as an array period, an incident angle and the like into a high-dimensional space, and then carrying out feature fusion, and the Transformer module comprises an encoder and a decoder structure, wherein the encoder comprises a multi-head self-attention module, a normalization layer and a feedforward neural network; the method comprises the steps of 1.2, establishing a global network, wherein the global network comprises a semantic and numerical feature fusion module and a Transformer module, firstly establishing a semantic information base according to the array unit position in the semantic and numerical feature fusion module, then converting space semantic information of each unit into a trainable vector based on the existing Wordembedding technology, modifying an initial current coefficient of each unit based on global coupling, firstly splicing the initial current coefficient with numerical features such as an array period, an incidence angle and a scale, and then fusing the initial current coefficient with the semantic vector in a high-dimensional space, and the Transformer module comprises an encoder and a decoder structure, wherein the encoder comprises a multi-head self-attention module, a normalization layer and a feedforward neural network, and the decoder comprises a mask attention module, a multi-head cross-attention module, a normalization layer and a feedforward neural network, acquiring a final current coefficient of a large-scale array after reasoning by the Transformer module, and acquiring final surface current distribution.
3. The intelligent sub-global basis function method for calculating electromagnetic scattering characteristics of a periodic object according to claim 1, wherein the implementation process of step 2 is as follows: Step 2.1, generating a 3X 3 subarray with different structural parameters and different excitations and a large-scale array sample, wherein for the 3X 3 subarray, the pitch angle and the azimuth angle of an incident wave are respectively set to be (theta, phi), the frequency of the incident wave is set, the sampling ranges of the theta and the phi are respectively set to be 0 DEG to 90 DEG, the sampling interval is respectively set to be 5 DEG and 10 DEG, the array period is respectively set to be (d x , d y ）, d x , d y lambda to 1.0 lambda, lambda is a wavelength, the sampling interval is 0.1 lambda) along the x-axis and the y-axis direction, and for the large-scale array, the array scale is set to be (N x , N y ）, N x ×N y is sequentially set to be 10X 10, 15X 15, 30X 30, wherein the setting range of the theta, phi and d x , d y of the large-scale array is the same as that of the subarray; and 2.2, preprocessing data, namely acquiring current coefficients of all array samples in the step 2.1 by adopting an existing NASED method as tag values of a sub-domain network and a global domain network, splitting the current coefficients into a real part and an imaginary part to be used as outputs of the sub-domain network and the global domain network together, setting numerical characteristics of the sub-domain network as (d x , d y , theta, phi) and numerical characteristics of the global domain network as (N x , N y , d x , d y , theta, phi), and carrying out normalization processing on the input characteristics and the tag values by adopting an existing Z-score method to reduce numerical difference.
4. A method for calculating intelligent sub-global basis functions of periodic target electromagnetic scattering properties according to claim 1 as described in claim 1, wherein the implementation process of step 3 is as follows: step 3.1, creating a mixed data and physical mixed loss function, wherein the mixed data and physical loss function consists of the numerical errors of the predicted value and the label value and the physical loss of the predicted value solving the impedance matrix equation, and the mixed loss function (which can be written as ; Wherein the method comprises the steps of And The outputs of the sub-domain network and the global network are the real part and the imaginary part of the current coefficient, thus the numerical loss function based on the mean square error is written as ; Where MSE represents the mean square error calculation, N is the total number of units used for training, alpha pre and alpha gt are the predicted value and the truth label of the current coefficient, respectively, alpha r i,pre and alpha im i,pre represent the real part and the imaginary part of the current coefficient of the i-th predicted NASED base function, respectively, alpha r i,gt and alpha im i,gt represent the real part and the imaginary part of the current coefficient of the i-th real NASED base function, respectively, K is the number of NASED base functions on each unit, Is L2 norm, physical loss writing based on impedance matrix equation ; Wherein R phy,j is the j-th physical residual term, M is the number of residual terms, Z RED represents the impedance matrix of the array, V r j and V im j represent the real and imaginary parts of the excitation matrix, respectively, d v is the dimension of V, and R and im represent the real and imaginary parts, respectively; Setting super parameters, namely setting the model dimension of a sub-domain network as 32, the learning rate as 0.001, the number of layers of an encoder as 3, the number of layers of a decoder as 2, and the batch size as 256, setting the model dimension of the global network as 64, the learning rate as 0.001, the number of layers of the encoder as 3, the number of layers of the decoder as 2, and the batch size as 1024; step 3.3, dividing a data set, namely using 80% of data in the data set for training, and using the rest data for testing a sub-domain network and a global network, wherein the training set is used for training the coupling modes in a large-scale finite period array for learning the sub-domain network and the global network; And 3.4, training a model, namely firstly training a sub-domain network by using a divided data set, respectively loading semantic information and numerical structure information in training set data into the sub-domain network model for forward propagation, calculating mixed loss of forward propagation each time, reversely calculating gradient based on the current mixed loss, updating network parameters by combining an existing Adam optimizer, testing by using a test set in each training, calculating training loss and testing loss, if the model is converged, saving the network structure and the network parameters of the model, then training a global network by combining initial current coefficients, and saving the model parameters after the network model is converged.
5. The method for calculating the intelligent sub-global basis function of the electromagnetic scattering property of the periodic object according to claim 1, wherein the implementation process of the step 4 is as follows: Step 4.1, obtaining NASED base functions, namely for a large-scale finite period array structure, firstly solving and analyzing a 3 multiplied by 3 subarray to obtain nine types of NASED base functions I 1 ,I 2 ,…,I 9 ; and 4.2, predicting the current coefficient by using a model, namely rapidly acquiring the initial current coefficient of the subarray by using a subarray network, and acquiring the initial current coefficient of a corresponding structure by only trimming the subarray network structure by combining a migration learning technology if the unit structure of the array changes. The initial current coefficients of nine units in the subarray and the structural characteristics of the target array are input into a global network together for coefficient correction, and finally the current coefficients of the target array are obtained; step 4.3, calculating a target current distribution the current distribution of the target array may be expressed as ; ; Where r represents the position vector of the array observation point, J array is the current distribution of the target array, N 0 is the total number of cells, K is the number of NASED basis functions on each cell, alpha n k is the kth NASED basis function on the nth cell, g n k (r) is the current coefficient of the kth NASED basis function on the nth cell, f n,m (r) is the mth RWG basis function of the nth cell, I n,m,k is the current coefficient of the mth RWG basis function of the nth cell, M is the number of RWG basis functions on each cell, and finally far field scattering is calculated according to the current distribution.

Description

Intelligent sub-global basis function method for calculating electromagnetic scattering characteristics of periodic target Technical Field The invention belongs to the field of computational electromagnetics, and particularly relates to an intelligent sub-universe basis function method for calculating periodic target electromagnetic scattering characteristics, which is used for rapid analysis and calculation of large-scale limited periodic array scattering characteristics. Background Large-scale finite period arrays are now widely used in communications base stations, phased array systems, super-surface military or civilian equipment, and the like. For many years, extensive research has been conducted on methods that accurately and effectively address the problem of large-scale limited-period array scattering. Unlike infinite periodic structures, large-scale finite periodic arrays need to take into account edge effects. With the expansion of array scale and the increasing complexity of cell structures, commercially available electromagnetic simulation software such as CST, HFSS, FEKO, etc. is faced with excessive computational memory requirements and high computational time costs. The traditional sub-global basis function method is time-consuming to fill and solve the impedance matrix equation even after the acceleration of combining the equivalent dipole and the fast dipole method, and especially, multiple parameter adjustments and optimizations are often required in the array design process, which causes the designer to spend high time cost in the simulation process. The combination of the traditional sub-global basis function method with a data-driven artificial neural network has been shown to be effective in analyzing scattering problems for large-scale finite-period arrays. However, the limited generalization capability and the high cost of the data set generation process are urgent challenges that data driven network models currently need to face. Disclosure of Invention The invention aims to accurately and efficiently analyze the electromagnetic characteristics of a large-scale limited periodic array, enhance the generalization capability of a network model and greatly reduce the high time cost for generating a data set. In order to achieve the above purpose, the invention provides an intelligent sub-global basis function method for calculating the electromagnetic scattering characteristics of a periodic target, which mainly comprises the following steps: The method comprises the steps of 1, establishing a sub-domain network and a global network, wherein the intelligent sub-domain base function method consists of the sub-domain network and the global network, the sub-domain network is used for predicting initial current coefficients of a 3 multiplied by 3 sub-array, and the global network corrects the initial current coefficients of each unit based on global coupling of a large-scale array; Step 2, data set making, namely establishing a data set according to physical characteristics of the 3 multiplied by 3 subarrays and the global arrays with different scales, generating labels by using a traditional NASED basis function method, and preprocessing the data; And step 3, training a model, namely taking a physical equation fused with an impedance matrix equation and a numerical error equation as a loss function, dividing a training set and a testing set according to a certain proportion, training and testing a network model to obtain optimal super parameters, and finally storing optimal model parameters. When the unit structure is changed, only the mobility learning fine adjustment sub-domain network is combined to quickly acquire initial current coefficients of different structures, and the global domain network does not need to be adjusted or retrained; Step 4, calculating far field scattering characteristics of the large-scale finite periodic structure, namely firstly solving the problem of the 3 multiplied by 3 subarrays to obtain NASED basis functions, then sequentially inputting the characteristics of each unit in the target structure into a trained sub-domain and global network to obtain final current coefficients, and finally calculating current distribution of the finite periodic array to calculate the far field scattering characteristics of the finite periodic structure; The specific implementation process of the step 1 is as follows: And 1.1, establishing a sub-domain network, wherein the sub-domain network comprises a semantic and numerical feature fusion module and a transducer module. Firstly, in a feature fusion module, the upper left corner, the upper right corner, the left edge, the inner center, the right edge, the lower left corner, the lower edge and the lower right corner of the array unit are used for constructing a semantic information base, then the spatial semantic information of each unit is converted into trainable vectors based on the prior Wordembedding technology, the sema